Remove the ‘Made with Google AI’ Sparkle and Label

Some images from Google’s Gemini models arrive with more than the small corner sparkle. They also carry a “Made with Google AI” text label, and often a set of provenance tags written into the file. If you created the image and want a clean version for your own use, this guide covers how to remove the visible parts and, more usefully, what is left behind once the pixels look clean.

The short version: the sparkle and the label are pixels you can edit, but the “Made with Google AI” claim also lives in places you cannot see. Understanding that split is the whole point, because it decides which tool you actually need.

#Two visible marks, two invisible ones

A single Gemini image can be marked in up to four ways. Two of them you can see and edit; two of them you cannot, and no amount of cropping changes them.

#Removing the sparkle and the text label

Both visible marks are handled the same way as any overlay: you either remove the pixels they occupy or reconstruct what should sit underneath. The label is often larger and lands over more of the image than the sparkle, so it usually needs reconstruction rather than a simple crop.

1. Crop when the sparkle and label sit in a margin you can afford to lose.
2. Reverse-alpha-blend on flat backgrounds, where the overlay can be mathematically subtracted for near pixel-perfect recovery.
3. Inpaint or regenerate when the mark covers texture, faces, or detail, then inspect the repaired area at full size.

A tool built for this, like our Gemini watermark remover, detects the sparkle and label automatically and reconstructs the region, so you are not masking pixels by hand. For a step-by-step on the corner mark specifically, see removing the Gemini (Nano Banana) watermark.

#The provenance you cannot see

This is where the “Made with Google AI” label gets deceptive. The words on the image are only one expression of that claim. The same provenance is frequently written into the file as C2PA Content Credentials or EXIF and XMP tags, and Gemini images also carry the imperceptible SynthID signal in their pixels. Erasing the on-screen text does not touch any of that.

This matters because a platform, client, or verification system rarely reads the visible label. It reads the metadata and runs a detector. So an image with the text painted out can still be identified as AI-generated. If your aim is to change what those systems report, our guide to image metadata and Content Credentials explains what actually travels with a file and why re-saving is not proof of removal.

#Verify before you trust the result

Whatever tool you use, confirm the outcome yourself rather than assuming a clean-looking image is a clean file.

• Check for the embedded signal with Google’s SynthID Detector or the Gemini app.
• Inspect provenance at contentcredentials.org/verify.
• Open the file’s metadata to see whether C2PA, EXIF, or XMP tags remain.
• Zoom to 100% on any region you inpainted to catch reconstruction artifacts.

#What Delete SynthID handles here

Delete SynthID addresses all three removable layers in one pass (the visible sparkle and label, the file metadata, and the embedded signal) across JPEG, PNG, WebP, AVIF, HEIC, and HEIF files up to 20MB. Your first image is free, so you can run the checks above before deciding whether it did what you need. It will not guarantee that every detector returns a clean result, because no honest tool can. If you want the fuller conceptual background first, start with what SynthID is or review pricing.